Post-Trained Nemotron Writes CUDA Kernels: KernelBench L2 22.8% to 82.1%
NVIDIAAI · x · 2026-10-17
NVIDIA shared Mindbeam's announcement of a collaboration to optimize open-source models: they post-trained Nemotron 3.5 Super VL to translate PyTorch modules directly into raw CUDA kernels.
- KernelBench Level 2 accuracy jumped from 22.8% to 82.1%, beating GPT 5.6 Sol's 57%
- The Litespark Inference Engine provides day-0 support, reaching 962 tokens/s per user on just two H200 GPUs — a 6.9x improvement over vLLM
- Mindbeam notes that generating performance-critical systems code is only half the battle; running it efficiently in production is the other
More from Models
- Hash Rate Ep. 189: Jev models on Bittensor subnet SN92, decentralized training vs traditional LLMs — markjeffrey · 2026-10-17
- 'Remove AI tics and mannered prose': a one-line prompt fix for Argon's writing — arjunrajlab · 2026-10-17
- Blind test of 200 real image edits: nano banana pro loses to its cheaper base model — ravisparikh · 2026-10-17
- Dev ditches Gemini 4 Argon for Grok 4.5, which eats 96.6% of his Cursor tokens — gaganghotra_ · 2026-10-17
- LMArena opens Code Arena WebDev for blind head-to-head model coding battles — arena · 2026-10-17
- Unverified: GPT-6 Pro said to launch parallel research subagents, 50-100 weekly messages — TheMoonMidas · 2026-10-17